Revert in_place exif_transpose, f16 loads

in_place requires a fairly recent version of pillow and performing the
performance benefit of it is minimal.

Loading as float16 does break other nodes so the change is getting
reverted.

Resolves #226

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "comfyui-videohelpersuite"
 description = "Nodes related to video workflows"
-version = "1.1.0"
+version = "1.1.1"
 license = "LICENSE"
 dependencies = ["opencv-python", "imageio-ffmpeg"]
 

videohelpersuite/load_images_nodes.py

@@ -49,7 +49,7 @@ def images_generator(directory: str, image_load_cap: int = 0, skip_first_images:
     for image_path in dir_files:
         i = Image.open(image_path)
         #exif_transpose can only ever rotate, but rotating can swap width/height
-        ImageOps.exif_transpose(i, in_place=True)
+        i = ImageOps.exif_transpose(i)
         has_alpha |= 'A' in i.getbands()
         count = sizes.get(i.size, 0)
         sizes[i.size] = count +1
@@ -59,9 +59,9 @@ def images_generator(directory: str, image_load_cap: int = 0, skip_first_images:
     iformat = "RGBA" if has_alpha else "RGB"
     def load_image(file_path):
         i = Image.open(file_path)
-        ImageOps.exif_transpose(i, in_place=True)
+        i = ImageOps.exif_transpose(i)
         i = i.convert(iformat)
-        i = np.array(i, dtype=np.float16)
+        i = np.array(i, dtype=np.float32)
         #This nonsense provides a nearly 50% speedup on my system
         torch.from_numpy(i).div_(255)
         if i.shape[0] != size[1] or i.shape[1] != size[0]:
@@ -104,13 +104,13 @@ def load_images(directory: str, image_load_cap: int = 0, skip_first_images: int
 
     if meta_batch is not None:
         gen = itertools.islice(gen, meta_batch.frames_per_batch)
-    images = torch.from_numpy(np.fromiter(gen, np.dtype((np.float16, (height, width, 3 + has_alpha)))))
+    images = torch.from_numpy(np.fromiter(gen, np.dtype((np.float32, (height, width, 3 + has_alpha)))))
     if has_alpha:
         #tensors are not continuous. Rewrite will be required if this is an issue
         masks = images[:,:,:,3]
         images = images[:,:,:,:3]
     else:
-        masks = torch.zeros((images.size(0), 64, 64), dtype=torch.float16, device="cpu")
+        masks = torch.zeros((images.size(0), 64, 64), dtype=torch.float32, device="cpu")
     if len(images) == 0:
         raise FileNotFoundError(f"No images could be loaded from directory '{directory}'.")
     return images, masks, images.size(0)

videohelpersuite/load_video_nodes.py

@@ -101,7 +101,7 @@ def cv_frame_generator(video, force_rate, frame_load_cap, skip_first_frames,
         frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
         # convert frame to comfyui's expected format
         # TODO: frame contains no exif information. Check if opencv2 has already applied
-        frame = np.array(frame, dtype=np.float16)
+        frame = np.array(frame, dtype=np.float32)
         torch.from_numpy(frame).div_(255)
         if prev_frame is not None:
             inp  = yield prev_frame
@@ -142,8 +142,8 @@ def load_video_cv(video: str, force_rate: int, force_size: str,
         #TODO: verify if garbage collection should be performed here.
         #leaves ~128 MB unreserved for safety
         memory_limit = (psutil.virtual_memory().available + psutil.swap_memory().free) - 2 ** 27
-    #space required to load as f16, exist as latent with wiggle room, decode to f32
-    max_loadable_frames = int(memory_limit//(width*height*3*(2+4+1/10)))
+    #space required to load as f32, exist as latent with wiggle room, decode to f32
+    max_loadable_frames = int(memory_limit//(width*height*3*(4+4+1/10)))
     if meta_batch is not None:
         if meta_batch.frames_per_batch > max_loadable_frames:
             raise RuntimeError(f"Meta Batch set to {meta_batch.frames_per_batch} frames but only {max_loadable_frames} can fit in memory")
@@ -153,7 +153,7 @@ def load_video_cv(video: str, force_rate: int, force_size: str,
         gen = itertools.islice(gen, max_loadable_frames)
 
     #Some minor wizardry to eliminate a copy and reduce max memory by a factor of ~2
-    images = torch.from_numpy(np.fromiter(gen, np.dtype((np.float16, (height, width, 3)))))
+    images = torch.from_numpy(np.fromiter(gen, np.dtype((np.float32, (height, width, 3)))))
     if meta_batch is None:
         try:
             next(original_gen)